$1.67 per professional task is Alibaba’s headline figure for Qwen4-Max, down from $4.43 for Qwen3.8-Max, with the new model keeping a 2.4-trillion-parameter MoE backbone and a 1-million-token context window.
35% fewer tokens in high-effort reasoning mode and lower list prices—$1.50 per million input tokens and $4.50 per million output tokens, versus $2 and $6—drove the claimed savings, while Alibaba also cited gains in coding and agentic benchmarks.
31.8% of functions passed both functionality and security tests on Last Secure Code Bench, up from 22%, after Alibaba said it filtered offensive cyber training data while retaining defensive material.
Greyhound Research and Security Brigade said buyers should treat the cost and safety claims as vendor benchmarks until external red teams or government evaluators verify them, especially because offensive and defensive cyber knowledge often overlaps.
30-day prompt retention, longer storage for flagged content, and planned open-weight smaller Qwen4 releases add governance and customization risks for enterprises, particularly after scrutiny of an earlier experimental Alibaba model’s unexpected network behavior.